arX

iv:1

606.

0934

3v3

[m

ath.

CO

] 8

Oct

201

9

A Variant of the Erdos-Sos Conjecture

Frederic Havet∗ Bruce Reed† Maya Stein‡ David R. Wood§

October 9, 2019

Abstract

A well-known conjecture of Erdos and Sos states that every graph withaverage degree exceedingm−1 contains every tree withm edges as a subgraph.We propose a variant of this conjecture, which states that every graph ofmaximum degree exceeding m and minimum degree at least ⌊2m3 ⌋ containsevery tree with m edges.

As evidence for our conjecture we show (i) for every m there is a g(m)such that the weakening of the conjecture obtained by replacing the first m

by g(m) holds, and (ii) there is a γ > 0 such that the weakening of theconjecture obtained by replacing ⌊2m3 ⌋ by (1− γ)m holds.

1 Introduction

A recurring topic in extremal graph theory is the use of degree conditions (suchas minimum/average degree bounds) on a graph to prove that it contains certainsubgraphs. For instance, every graph of minimum degree exceeding m− 1 containsa copy of each tree with m edges. (Embed the root anywhere, and greedily continueembedding vertices whose parents are already embedded.)

∗CNRS, Projet COATI, I3S (CNRS and UNS) UMR7271 and INRIA, Sophia An-tipolis, France. Research supported by ANR under contract STINT ANR-13-BS02-0007(Frederic.Havet@cnrs.fr).

†School of Computer Science, McGill University, Montreal, Canada (breed@cs.mcgill.ca).CNRS, Projet COATI, I3S (CNRS and UNS) UMR7271 and INRIA, Sophia Antipolis, France(reed@i3s.unice.fr). Visiting Research Professor, ERATO Kawarabayashi Large Graph Project,Japan.

‡Department of Mathematical Engineering and Center for Mathematical Modeling (UMI 2807CNRS), Universidad de Chile, Santiago, Chile (mstein@dim.uchile.cl). Research supported byCONICYT + PIA/Apoyo a centros cientıficos y tecnologicos de excelencia con financiamientoBasal, Codigo AFB170001, by FONDECYT Regular Grant 1183080, and by Millennium NucleusInformation and Coordination in Networks.

§School of Mathematics, Monash University, Melbourne, Australia (david.wood@monash.edu).Research supported by the Australian Research Council.

1

In 1963, Erdos and Sos conjectured the following strengthening of this fact: ifa graph has average degree exceeding m − 1 then it contains every tree with medges as a subgraph. Their conjecture has attracted a fair amount of attentionover the last decades. Partial solutions are given in [BD96, Hax01, SW97], and inthe early 1990’s, Ajtai, Komlos, Simonovits and Szemeredi announced a proof ofthis result for sufficiently large m. In order to see that the Erdos-Sos conjectureis best possible, observe that no (m− 1)-regular graph contains the star K1,m as asubgraph. Alternatively, consider a graph that consists of several disjoint copies ofthe complete graph Km; this graph contains no tree with m edges as a subgraph.

The related Loebl-Komlos-Sos conjecture from 1995 [EFLS95] states that if agraph has median degree at least m then it contains every tree with m edges as asubgraph. This conjecture had also received considerable attention [AKS95, Coo09,HP15, PS12, Zha11], and recently, an approximate version was shown in [HKP+17a,HKP+17b, HKP+17c, HKP+17d] (see also [HPS+15]). Note that the examples abovedemonstrate that the Loebl-Komlos-Sos conjecture is tight as well.

In this paper we propose a new conjecture for tree embeddings under degreeassumptions. We consider the minimum and maximum degrees rather than theaverage or median degrees.

Conjecture 1.1 If a graph has maximum degree at least m and minimum degreeat least ⌊ 2m

3⌋ then it contains every tree with m edges as a subgraph.

We remark that every graph of average degree exceeding m− 1 has a subgraphof minimum degree at least m

2. Thus, replacing ⌊2m

3⌋ by m

2in our conjecture would

give a strengthening of the Erdos-Sos conjecture. However, two simple examplesshow that the value ⌊2m

3⌋ here is best possible. In both examples we consider the

tree T with 3k+ 1 vertices obtained from three stars on k vertices by adding a newvertex v adjacent to their centers. In the first example, G is the graph obtainedfrom two copies of K2k−1 by adding a universal vertex. In the second example, G isthe graph obtained by adding a universal vertex to K2k−2,2k−2.

Nevertheless, focussing on the minimum degree of the graphs in question, couldbe an effective technique for approaching the Erdos-Sos conjecture. Indeed, it mightbe possible to prove a natural common generalization of Conjecture 1.1 and theErdos-Sos Conjecture which makes no mention of the average degree.

Note that Conjecture 1.1 holds for paths (even with the weaker bound of m2

on the minimum degree), because of the well-known Dirac-type result that everyconnected graph G of minimum degree δ(G) has a path on min{2δ(G) + 1, |V (G)|}vertices. It also holds for trees with many leaves (see below).

As further evidence for Conjecture 1.1, we prove the following two weakenings.

Theorem 1.2 There is a function g such that if a graph has maximum degree atleast g(m) and minimum degree at least ⌊2m

3⌋ then it contains every tree T with m

edges as a subgraph.

2

Theorem 1.3 There is a γ > 0 such that if a graph has maximum degree at leastm and minimum degree at least (1−γ)m then it contains every tree T with m edgesas a subgraph.

After proving some useful results on trees in Section 2, we prove Theorem 1.2 inSection 3 and Theorem 1.3 in Section 4. While the proof of the first theorem is notvery hard, the proof of the second theorem is much more complicated. We dedicatethe remainder of the introduction to a sketch of some of the ideas used in both ourproofs. For a more detailed sketch of the proof of Theorem 1.3 we refer the readerto the beginning of Section 4.

Let us start with an easy observation that involves trees having a vertex s thatis adjacent to many leaves. We can embed s in a maximum-degree vertex f(s) ofthe host graph, and then embed the rest of the tree, except for the leaves adjacentto s, in a greedy fashion. Finally, we embed the leaves at s, exploiting the largedegree of f(s). Note that this procedure gives a proof of both our theorems, and ofConjecture 1.1, for all trees that contain a vertex adjacent to at least ⌈m

3⌉ leaves. It

also proves Theorem 1.3 for all trees that contain a vertex adjacent to at least ⌈γm⌉leaves. In particular, this proves the conjecture and Theorem 1.2 for trees having avertex of degree at least ⌈2m

3⌉, and Theorem 1.3 for trees having a vertex of degree

at least ⌈ (1+γ)m2

⌉.The proof of both of our theorems for the remaining trees splits into two cases

depending on whether or not the host graph G has a small dense subgraph. Toilluminate why small dense subgraphs are important let us now prove the conjecturefor host graphs which do not contain any connected subgraphs with m+ 1 verticeshaving average degree at least 2, that is, host graphs of girth at least m+ 2. If wegreedily embed a tree with m edges in such a graph by embedding the vertices inbreadth-first order, treating all the children of each vertex as a consecutive block,we see that for every non-root vertex s we have embedded, the girth conditionensures that its image f(s) is adjacent to the image of exactly one vertex of thetree (namely the parent of s). Since s has degree at most ⌊2m

3⌋, we will be able to

embed its children into the unoccupied neighbours of f(s). So the greedy embeddingstrategy succeeds in graphs of girth at least m+ 2.

Without the girth condition imposed in the illustrating example in the previousparagraph, but still assuming that our graph is relatively sparse and has no densesubgraphs (this is the first of the two cases mentioned above), we can still show thatonly a few vertices have many occupied neighbours. Our approach in this case isto try and stay away from these vertices when embedding the rest of the graph. Inorder to do so, we exploit the well known fact (see Section 2) that every tree T withm edges has a vertex z such that at most one component of T − z has more than athird of the vertices of T , and if such a component exists, it has at most two thirdsof the vertices of T . The same is true replacing ‘a third’ with γ and ‘two thirds’

3

with 1 − γ. This means that we can split the components of T − z convenientlyinto two sets, such that the one containing more vertices can be embedded greedilyusing the minimum degree of the host graph, while embedding z into a maximumdegree vertex. Now, for embedding the remaining vertices we need to stay awayfrom the occupied vertices. In proving Theorem 1.2 this is relatively easy to dobecause f(z) has huge degree, and so we have a lot of flexibility when placing theneighbours of z. In proving Theorem 1.3, f(z) may only have m neighbours whichmakes things harder. In this case we need to be more careful during the first phaseof the embedding. Here, the higher minimum degree comes in handy.

Turning to graphs with small dense subgraphs (the second case mentioned above),we only discuss the proof of Theorem 1.3 here, as the approach taken in the proofof Theorem 1.2 is fairly straightforward. In that proof, we focus on the densestsubgraphs of the host graph with at most m+1 vertices. For every such maximum-density subgraph H , if H has minimum degree d then every vertex outside of H seesat most d+1 vertices of H . Furthermore, because H is small and dense it turns outthat we can embed trees with significantly more than d vertices in H .

So we can often embed significantly more than d+ εm vertices of the tree in Hand just slightly less than (1−ε)m−d in G−H which has minimum degree at least(1− ε)m− d− 1. In order to do so, we split the tree T , by determining a cutvertexz, and grouping the components of T − z into two sets of components, C1 and C2,whose sizes fit well with our embedding plans.

There are some further complications: we need to consider some extensions ofthese small dense graphs, some dense bipartite graphs, and a partition of the graphinto such dense pieces. For more on these difficulties, see Section 4. We hope ourdescription here is enough to give a flavour of the proof.

Finally, we mention that recent work [BPMS19, RS19a, RS19b] has partiallyconfirmed Conjecture 1.1.

2 Some Properties of Trees

In this section we prove some useful results on trees. Our first aim is to find arelatively large stable set whose vertices have degree at most 2 in the tree. Thesmall degree of the vertices in this set means that when embedding into a smalldense subgraph H , we will be able to embed them last, after (carefully) embeddingthe rest of the vertices, thereby embedding many more than δ(H) vertices into H .

Lemma 2.1 Every rooted tree T with at least two vertices contains a stable set ST

of size ⌈|V (T )|/6⌉ not containing the root such that:

(a) every vertex in ST is a leaf, or a vertex of degree 2 whose parent is also a non-rootvertex of degree 2, and

(b) no child of a vertex in ST is the parent of some other vertex of ST .

4

Proof. Letting ℓ be the number of non-root leaves of T , we see that removingthe root of T and all vertices of degree greater than 2 in T splits the non-root verticesof degree 1 and 2 in T into fewer than 2ℓ paths of with total number of verticesat least |V (T )| − ℓ. We can put every third vertex within each of these paths intoST , as long as we start with the second from the root. We can thereby ensure that|ST | > (|V (T )| − 3ℓ)/3. On the other hand, we can simply put all the non-rootleaves of T into ST , so |ST | > ℓ. The result follows.

Also, it turns out that matchings in the tree we wish to embed can be useful whenembedding into a dense subgraph. This is because we can embed matched verticesone right after the other, that way their embedding happens under almost identicalcircumstances (with respect to the ‘used’ or ‘unused’ parts of the host graph). Inaddition, for the first vertex of a matching edge we can choose an image with highdegree into some set we wish to use for the second vertex.

Lemma 2.2 For every tree T , and every 1 6 ℓ 6 |V (T )|/2, either T contains atleast |V (T )| − 2ℓ+ 2 leaves, or for every vertex v of T , there is a subtree of T with2ℓ vertices which contains v and has a perfect matching.

Proof. Consider a maximum subtree T ′ of T containing v which has a perfectmatching. If some component of T − T ′ has at least two vertices, then adding twoadjacent vertices of this component to T ′, including the one joined to T ′ by an edge,contradicts the maximality of T ′. So all vertices in V (T − T ′) are leaves, and since|V (T ′)| is even, the result follows.

Finally, we prove a much used observation that allows us to split the tree intosubtrees whose sizes we can control.

Observation 2.3 Let T be a tree on t vertices.

(a) There is z ∈ V (T ) such that every component of T −z has t/2 or fewer vertices.

(b) For any t′ < t2, either every component of T − z has fewer than t′ vertices or

there is a vertex vt′ of T − z such that the component of T −vt′ containing z hasat most t− t′ vertices and every other component has fewer than t′ vertices.

Proof. For (a), we root the tree and let z be the vertex furthest from the rootsuch that the subtree formed by z and its descendants contains at least half thevertices.

For (b), we can assume there is a component C1 of T − z having at least t′

vertices. We root the tree at z and let vt′ be the vertex furthest from z in C1 suchthat the subtree formed by vt′ and its descendants contains at least t′ vertices.

A separator for a tree T on t vertices is a vertex z such that each component ofT − z has at most t/2 vertices. Note that the choice of such a z is unique or thereare two such choices which are endpoints of an edge e such that each component ofT − e contains t/2 vertices.

5

3 The Proof of Theorem 1.2

Define g(m) := (m+ 1)2m+6 + 1 and consider a counterexample (m,G) minimizing|E(G)|. Let v be a vertex of G of degree g(m) and note that minimality impliesthat if uw is an edge of G− v then one of u or w has degree ⌊2m

3⌋. Let t = m + 1

and let t′ = m− ⌊2m3⌋. We assume t > 3 (otherwise the proof is easy). We split the

proof into two cases as follows.

Case 1: G contains a Kt3,t′.

Let A be the smaller side of this complete bipartite graph and let B be the largerside. Minimality implies that every vertex in B has degree ⌊2m

3⌋.

Thus, for any vertex b of B, there are fewer than t2 vertices of B which areadjacent to b or have a common neighbour with b which has degree at most t. So,we can choose a stable subset B′ of t vertices of B, such that no vertex of degreeless than t sees two vertices of B′.

We take any subtree of T with 2t′+1 vertices and embed it in A∪B′ using morevertices of B′ than of A. We can now greedily complete the embedding, since bythe choice of B′, every (used or unused) vertex of degree less than t has degree atleast ⌊2m/3⌋−1 = m− (t′ +1) into G−B′, while at least t′+1 vertices are alreadyembedded into B′.

Case 2: G contains no Kt3,t′ .

Note that in this case, for any subset S of V that contains at least t′ vertices,we have

less than t3(|S|

t′

)

vertices of G see t′ or more vertices of S. (1)

Applying Observation 2.3 with our chosen value of t′, we see that there is avertex w of T such that no component of T −w has more than 2m

3vertices, and all

but the largest component have fewer than t′ vertices.We embed w into v. We greedily embed the largest component of T −w into G.

We then embed the remaining components of T −w, which have size at most t′ − 1.Whenever we come to embed such a component K, we proceed as follows.

Let A0 be the set of vertices into which we have already embedded a vertex of T(before starting to embed K). Successively, for i = 1, . . . , t′, let Ai ⊆ V (G)−⋃

j<iAj

consist of all those vertices that have degree less than t′ − 1 in G −⋃

j<iAj . Notethat A0, . . . , At′ are pairwise disjoint.

Now, each vertex of A1 has degree at least ⌊2m/3⌋− (t′−2) > t′ into A0. Hence,we can use (1) to see that

|A1| 6 t3(|A0|

t′

)

.

6

For i > 2, note that if v ∈ Ai, then (since v /∈ Ai−1), we know that v has a neighbourin Ai−1. So for i > 2, the definition of Ai−1 gives that |Ai| 6 (t′−2)|Ai−1|. Therefore,

|t′⋃

i=0

Ai| 6t′∑

i=0

(t′ − 2)it3(|A0|

t′

)

6 t2t+4.

So, since g(m) > t2t+4, there is a neighbour of v outside of⋃t′

i=0Ai in whichwe can embed the neighbour x of w in K. We now use the degree condition onthe sets Ai to greedily embed K levelwise, allowing each level j (that is, the jth

neighbourhood of x) to use vertices in G−⋃t′−ji=0 Ai. This way we ensure that A0 is

not used for our embedding of K.Iterating this process for each yet unembedded component of T − w proves

Theorem 1.2.

4 The Proof of Theorem 1.3

Let us start by giving an overview of our proof. Our proof has five parts. In the firstpart, in Subsection 4.1, we show that if all the subgraphs of the host graph G withat most m + 1 vertices are really sparse, then we can find the desired embedding(this is done in Lemma 4.1). Thus we can assume that G has a subgraph of size atmost m+ 1 that is reasonably dense, i.e. has average degree linear in m.

In Subsection 4.4, we show how to use such a subgraph. If we cannot find thedesired embedding of T , then we find a very dense subgraph H of G. More precisely,either H has at most m + 1 vertices, and is almost complete, in the sense that atevery vertex there is only a small fraction of the possible edges missing, or H isalmost complete bipartite (in the same sense), with each of its sides having size atmost m.

Such a subgraph H can be very useful for embedding a part of the tree, asits extreme density allows us to accommodate more vertices of T than we wouldexpect if we were only using the minimum degree. For technical reasons, it willbe convenient to explain this approach in detail already in Subsections 4.2 and 4.3(before actually finding H in Subsection 4.4). A series of lemmas given in these twosubsections covers the range of possible situations that we might have to deal within a later stage of the proof, when we wish to embed parts of T into such a graphH .

In the last part, in Subsection 4.5, we put everything together. We find a maximalset of disjoint very dense subgraphs Hi, knowing that at least one such subgraphis guaranteed to exist by what we said above. (Actually, our Hi will be slightexpansions of the subgraphs found in Subsection 4.4.) We show that if we cannotembed T using the results of Subsections 4.2 and 4.3, there are only very few edgesbetween the different subgraphs Hi, and between the union of theHi and the leftover

7

of the graph G. By the results of Subsection 4.1, this means that there is no suchleftover (as it would have to contain another very dense graph). Thus one of the Hi

contains a vertex of degree at least m. Again making use of results of Subsections 4.2and 4.3, we show we can embed T . This completes the overview of our proof.

We close this subsection with some preliminaries. We often iteratively constructan embedding f of T in such a way that the embedded subtree is always connected.In this case, whenever we come to embed a vertex s of T , there is a unique embeddedneighbour p(s) of s and we need only ensure that s is embedded in a neighbour off(p(s)) which has not yet been used in the embedding. We refer to this as a gooditerative construction process.

Note that we can and do assume that no vertex of T is incident to more thanγm leaves, as otherwise we can simply embed this vertex into a maximum degreevertex, greedily embed the tree except for the leaves incident to it and then embedthese leaves. For this reason, all our lemmas are stated with this assumption.

4.1 Locally Sparse Graphs

A graph is locally m-sparse if it contains no subgraph with at most m + 1 verticesand average degree exceeding m

25. The main result of this section is the following:

Lemma 4.1 Suppose T is a tree with at most m edges each of whose vertices isadjacent to at most m

20leaves and G is a locally m-sparse graph of minimum degree

at least 19m20

. Then for any vertex w of G and separator z for T , we can find anembedding f of T in G such that f(z) = w.

Once we have proved Lemma 4.1, we can continue our proof only consideringhost graphs G that are not locally m-sparse. Before proving Lemma 4.1, we showan auxiliary result.

Lemma 4.2 Let G be a locally m-sparse graph of minimum degree at least 1920m.

Then for any S ⊆ V (G) with |S| 6 m− 1 there is a set S ′ ⊇ S such that G−S ′ hasminimum degree at least m

2and |S ′| 6 |S|+ m

20.

Proof. Assume for a contradiction that there is no such set S ′. For any setS ′ ⊇ S such that |S ′| 6 |S| + m

20, there is a vertex a of G − S ′ having degree less

than m2in G− S ′, and so more than δ(G)− m

2= 9m

20neighbours in S ′. In particular

|S| > 9m20, and we can find a set A with ⌈m

20⌉ vertices such that every vertex in A has

at least 9m20

neighbours in S ∪ A. (Find A by successively adding suitable vertices).We choose any set B ⊆ S of size |A| − 1, and note that the set (S − B) ∪ A has atmost m vertices and induces more than (9m

20− |B|)|A| > 8m2

400= m2

50edges, and thus,

its average degree is above m25. This contradicts G being locally m-sparse.

8

Proof of Lemma 4.1. We let F be the union of some of the components ofT − z which together have between m

4and m

2vertices. If we can embed T −F into a

set f(V (T −F )) that avoids at least |F | − 1+ m20

neighbours of w, then by applyingLemma 4.2 to S0 = f(V (T − F )), we obtain a set S ′

0 that avoids at least |F | − 1neighbours of w and such that G − S ′

0 has minimum degree m2. Now z is adjacent

to at most m20

leaves it is adjacent to at most |F | − 1 vertices of F , and so we canembed all neighbours of z in F into N(w) − S ′

0. Then, since |F | 6 m/2, we canextend this embedding greedily into an embedding of all of T . Hence, fixing any setN ⊆ N(w) with |N | =

⌈

19m20

⌉

, it suffices to embed T − F using at least m10

verticesoutside N .

Choose any set S ⊆ N+w containing w of size ⌈2m3⌉, and consider the set S ′ ⊇ S

given by Lemma 4.2. Then |S ′ − w| 6 3m4

and the vertices outside S ′ have degreeat least m

2into G − S ′. We now embed into N − S ′ either all or |N − S ′| of the

neighbours of z in T−F , and then embed the corresponding components of T−F−zgreedily, trying to avoid S ′ as much as possible. (Since |T −F − z| 6 3m

46

19m20

, wecan clearly embed all of these components in G.)

If we never used any vertex in S ′, then, depending on how many neighbours ofz we embedded into N − S ′, we embedded either all of T − F − z in G− S ′, or atleast |N − S ′| − m

20>

m10

vertices of T − F − z in G − N (since by assumption, zhas at most m

20leaf neighbours). Since z is not in N , in the first case we found the

desired embedding, and in the second case we can greedily continue to find it.So assume we used S ′, and let x be the first vertex we embedded there. Then,

since we tried to avoid S ′, the parent of x is embedded in a vertex that has at leastm2neighbours in G − S ′ that have already been used for our embedding. At least

m2− |N − S ′| of these vertices are outside N . But |N − S ′| 6 |N − S| 6 19m

20− 2m

3.

Hence at least m2− 19m

20+ 2m

3> m

10vertices outside N have been used for embedding

T − F , which is as desired. We greedily continue our embedding of T − F .

4.2 Filling Small Almost Complete Subgraphs

We now prove some auxiliary lemmas which are at the heart of the whole proof.They allow us to use almost complete subgraphs H ′ of the host graph G in orderto embed some suitable subtree T ′ of T . The point of these lemmas is that T ′ isallowed to be substantially larger than the minimum degree of H ′.

Lemma 4.3 Let 0 < ε < 1200

, let H ′ be a graph with minimum degree at least(1−2ε)(|V (H ′)|−1), and let T ′ be a tree with m′ edges, rooted at z, with (|V (H ′)|−1)/2 6 m′ 6 (1− ε)(|V (H ′)| − 1). If each vertex of T ′ is incident to at most εm′/2leaves, then we can embed T ′ in H ′, choosing any vertex as the image of z.

Proof. Using Lemma 2.1, we find a stable set S in T ′ of ⌈m′+16

⌉ non-root verticeswhich are leaves or vertices of degree 2 whose parents are non-root vertices of degree2, such that no child of a vertex in S is the parent of some other vertex of S. By

9

the definition of S, for any vertex in T ′ − S that has more than one child in S, allits children in S are leaves. Hence

each vertex in T ′ − S has at most εm′/2 children in S. (2)

So, we can choose a set S ′ ⊆ S with |S|/2 6 |S ′| 6 |S|/2 + εm′/2 such that novertex of S − S ′ is closer to z than any vertex of S ′, and such that for any givenvertex in T ′, either all or none of its children in S belong to S ′.

Since our assumption on ε ensures that

|S − S ′| > m′ + 1

12− εm′

2> 2ε(|V (H ′)| − 1),

the minimum degree of H ′ is large enough to allow us to use a good iterative con-struction process to greedily embed the component of T ′ −N(S − S ′) that containsz and all of S ′. (In particular, children of vertices in S ′ are embedded, but verticesfrom S−S ′ and their parents are not.) We immediately unembed the vertices of S ′.Note that |T ′ − S ′| 6 (1− 2ε)(|V (H ′)| − 1), so it is possible to greedily embed theremainder of T ′ − S ′. However, our plan is to embed the remainder of T ′ − S ′, in away that the vertices of S ′ can be embedded afterwards. So we do it cautiously.

Call a vertex u ∈ V (H ′) good for a vertex s ∈ S ′, if u is adjacent to both of theimages of the two neighbours of s in T ′. Let Bad be the set of all vertices u ∈ V (H ′)with the property that

there are more than|S ′|2

vertices in S ′ for which u is not good.

Since for each vertex s ∈ S ′ there are at most 4ε(|V (H ′)|−1) vertices u ∈ V (H ′)that are not good for s, it follows that there are at most 4ε(|V (H ′)| − 1)|S ′| pairss ∈ S ′, u ∈ V (H ′) such that u is not good for s. Therefore,

|Bad| 6 8ε(|V (H ′)| − 1). (3)

We now proceed our embedding of T ′ − S ′ in the following manner. When weare about to embed any vertex p which has one or more children in S − S ′, we tryto embed p into a vertex with many unused neighbours in Bad. Note that sinceeach unused vertex of Bad has at most 2ε(|V (H ′)| − 1) non-neighbours, there areat most 4ε(|V (H ′)| − 1) vertices which see less than half of the unused vertices ofBad. So, as the image of the parent of p has more than

(1− 2ε)(|V (H ′)| − 1)− |V (T ′)− {p} − S ′| > −ε(|V (H ′)| − 1) + |S ′|

>m′ + 1

12− ε(|V (H ′)| − 1)

> 4ε(|V (H ′)| − 1)

10

unused neighbours (here we use our upper bounds on ε andm′), we can embed p intoa vertex which sees more than half of the unused vertices of Bad. We immediatelyembed all children of p in S trying to embed as many as possible into unused verticesof Bad. By (2), we will be able to embed all these children, unless Bad has less thanεm′ unused vertices. Hence as long as Bad has less than εm′ unused vertices, weembed the vertices of S − S ′ in Bad. But, since |S − S ′| > m′

12− εm′

2> |Bad| − εm′,

eventually Bad will have less than εm′ unused vertices. After that, we embed allthe vertices greedily. Doing so, when we finish the embedding of T ′ − S ′, we haveused up all but at most εm′ vertices of Bad.

It remains to embed S ′. Consider the auxiliary bipartite graph between S ′ andthe set U of the so far unused vertices in H ′, i.e. the graph that has an edge su fors ∈ S ′, u ∈ U , if u is good for s. By Hall’s theorem, if we cannot embed S ′ in H ′,then in the auxiliary graph there is a (non-empty) set W ⊆ S ′ whose neighbourhoodis smaller than |W |. In other words, there is a subset UW ⊆ U such that |UW | < |W |,and no vertex in U − UW is good for any vertex in W .

Because of our assumption on the minimum degree of H ′, we know that |U −UW | 6 4ε(|V (H ′)|−1). On the other hand, by the other assumptions of the lemma,

|U | > |S ′|+ ε(|V (H ′)| − 1), (4)

and thus

|S ′ −W | < |S ′| − |UW | 6 |S ′| − |U |+ 4ε(|V (H ′)| − 1) 6 3ε(|V (H ′)| − 1).

So |W | > |S ′|/2 (because |S ′| > m′

12> 6ε(|V (H ′)|−1)), and therefore, U−UW ⊆ Bad.

Since U contains at most εm′ < ε(|V (H ′)| − 1) vertices of Bad (as we used all othervertices of Bad earlier), we deduce from (4) that |S ′| < |UW | < |W |, a contradiction.So we can embed all of S ′ as planned.

Observe that in the previous proof, we could have embedded an even larger treeT ′ in H ′, if we knew that the set Bad could be filled up completely during themiddle stage of the embedding, when we try to put as many vertices of S − S ′ aspossible into Bad. In fact, the term ε(|V (H ′)| − 1) from (4) (which comes fromthe assumption that m′ 6 (1 − ε)(|V (H ′)| − 1)) is only needed to make up for theunused vertices of Bad, in the inequality of the second-to-last line of the proof.

Under certain circumstances, we can fill up Bad completely, or up to a very smallfraction. This is the content of the next two lemmas.

Lemma 4.4 Let 0 < ε < 1200

, let H ′ be a graph with m′ + 1 vertices of minimumdegree at least (1−2ε)m′, and let v be a vertex of H ′ which sees all of V (H ′)− v. IfT ′ is a tree with at most m′ edges such that each vertex of T ′ is incident to at mostεm′/2 leaves, then we can embed T ′ in H ′.

11

Proof. Clearly we can assume T ′ is not a single vertex, so εm′

2> 1. We

repeatedly subdivide an edge from a leaf until T ′ has exactly m′ edges. Clearly, itis enough to prove the result for such T ′.

We will proceed very much as in the proof of Lemma 4.3, with two small dif-ferences. Firstly, we avoid v in our embedding throughout the process. As before,we stop right before reaching the parents of vertices in S − S ′, and then unembedthe vertices from S ′. We define the set Bad as in the proof of Lemma 4.3, andobserve that |Bad| 6 8ε|V (H ′)|. The next step is a little different from the proof ofLemma 4.3: When embedding the rest of T ′−S ′, every time we consider the parentof a vertex in S−S ′ we are happy if we embed at least half of its children in verticesof Bad. Since we always embed in a vertex which sees half of Bad, if we fail, thenthe current parent p has more children in S − S ′ than there are vertices in Bad. Inthis case, we embed p in v (this is possible as v sees all of V (H ′)− v), and use upall the vertices of Bad for embedding the children of p in S − S ′. Observe that weare bound to find such a vertex p, because |S−S′|

2> |Bad|. We then embed the rest

of T ′ − S ′ greedily.Now continue as in the proof of Lemma 4.3, and embed S ′. Note that although

in (4), we only get |U | > |S ′| instead of |U | > |S ′| + ε|V (H ′)|, we compensate forthis shortcoming by having filled up all of Bad. Namely, from U − UW ⊆ Bad wecan deduce that U = UW , and thus obtain |S ′| = |UW | < |W |, a contradiction whichshows that we can embed all of S ′ as planned.

The next lemma goes one step further than the previous lemmas, embedding thetree in- and outside the dense subgraph.

Lemma 4.5 For sufficiently small positive γ the following holds. Let T be a treewith m edges none of whose vertices is incident to more than γm leaves. Let H ′

be a subgraph of G with at most m + 1 + 3γm vertices such that (i) both H ′ andG− H ′ have minimum degree at least m − 3γm, and (ii) there is a vertex v of H ′

with degree at least m in G. Then we can embed T in G.

Proof. We can assume that v does not see m vertices of H ′, as otherwise we aredone by applying Lemma 4.4 to N(v) ∩H ′. We let a = m− |N(v) ∩H ′|, and notethat v has at least a > 1 neighbours outside of H ′. We embed a separator z for Tinto v.

If the sum s of the sizes of the a largest components of T − z is at least 3γm,or if T − z has less than a components, then we can choose some subset of thesecomponents that has between 3γm and m

2vertices. We embed these components

greedily in G−H ′, putting neighbours of z into neighbours of v, and then embed therest of T greedily in H ′ (note we can do so because of condition (i) of the lemma),and are done.

So assume from now on that T − z has at least a components and that

s 6 3γm. (5)

12

Letting F be the union of the a largest components, all the components ofT − F − z have size at most s

a. In particular,

no vertex of T − F other than z has degree exceedings

a. (6)

Also note thats > 2a, (7)

since there are at most γm singleton components of T − z, and by (5), these cannotbe part of F .

We embed F into G−H ′ and proceed as in the proof of Lemma 4.4, to embedT −F into H ′∩N(v) with one important difference, which we explain momentarily.

As before, we stop right before reaching the parents of the vertices in S − S ′,and then unembed the vertices from S ′. We define the set Bad as in the proof ofLemma 4.4, and observe that |Bad| 6 24γm. We continue embedding the rest ofT ′ − S ′, and as in Lemma 4.4, every time we consider the parent p of a vertex in Swe are happy if we embed at least half of its children in vertices of Bad. Let us callsuch a parent p a happy parent. Since we always embed in a vertex which sees halfof the unused vertices of Bad, if we cannot embed at least half of the children of pin Bad, then we can use up half the currently unused vertices of Bad by embeddingchildren of p. Let us call such a parent p an unhappy parent.

Next, we determine the size of the set of unused vertices of Bad at the end ofthis process. Observe that at least half of the vertices of S − S ′ with happy parentsget embedded in Bad, and thus, at most 2|Bad| vertices of S − S ′ can have happy

parents. So, at least m(1−3γ)12

− γm − 2|Bad| > m15

vertices of S − S ′ have unhappyparents. Thus, by (6) there are at least am

15sunhappy parents.

So, setting r = ms, we see that the number of unused vertices of Bad at the end

of the process is at most 24γrs2−ra15 . Since a > 1, and, by (5), r is at least 1

3γ, if γ

is sufficiently small then there are at most s4unused vertices of Bad left.

Now, note we are only embedding |T − z − F | = m− s vertices into N(v) ∩H ′,and the size of N(v) ∩ H ′ is at least m − a, which by (7) is at least m − s

2. This

means we have more vertices in which to embed than vertices we need to embedeven if we throw the unused vertices of Bad away. So, we can continue as in theproof of Lemma 4.3, and embed S ′.

4.3 Filling Small Almost Complete Bipartite Subgraphs

This section has a similar aim as the previous section. Instead of small almostcomplete subgraphs, we now focus on small almost complete bipartite subgraphs ofthe host graph G.

We chose to start this subsection with the following lemma, because of the strongsimilarities of its proof with the proofs from the previous subsection.

13

Lemma 4.6 Let 0 < ε < 1200

, let T ′ be a tree with m′ edges such that each vertex

of T ′ has at most εm′

2leaf children. Let (C,D) be the unique 2-colouring of T ′ with

|C| 6 |D|. Let H ′ = ((A,B), E) be a bipartite graph of minimum degree at least(1− 3ε)m′ such that both A and B have at most ⌊(1+ ε)m′⌋ vertices, B has at least|D| vertices, and A contains a vertex v which sees all of B. Then we can embed T ′

in H ′.

Proof. We can assume that m′ >2εor the tree must be a singleton and we are

done. Because of the minimum degree condition on H ′, we can greedily embed T ′

unless |C| < 3εm′+1 6 4εm′, so we assume this is the case. This implies that thereare at least (1−8ε)m′ leaves of T ′ in D (for this, observe that rooting T ′ arbitrarily,every non-leaf vertex in D has at least one child in C). We let T ′′ be T ′ with theseleaves removed. Our plan is to embed C in A and D in B, starting with T ′′.

We use a good embedding algorithm to begin embedding T ′′ in H ′ − v, startingwith a vertex of C. We pause the procedure the first time that the set X of verticesembedded in A has edges to more than half of the vertices of T ′ − T ′′. We let L bethe set of neighbours of X in T ′ − T ′′. Note that (1

2− 4ε)m′ 6 |L| < (1

2+ 1

2ε)m′, by

our assumption on the number of leaf children at each vertex.Let f(X) be the image of X . We assign each vertex x of X a weight wx which

is the number of vertices of L it is incident to. For every X ′ ⊆ X , we set w(X ′) =∑

x∈X′ wx. Note that w(X) = |L|. Call a vertex b ∈ B bad if there is a set X ′ ⊆ X

with w(X ′) > |L|2

such that b has no neighbour in f(X ′). We let Bad be the set ofall bad vertices of B.

We claim that Bad contains at most 8εm′ vertices. Indeed, otherwise everyvertex from f(X) sees more than half the vertices of Bad. Consider the graph weobtain from blowing up each of the vertices f(x) ∈ f(X) to a set f ′(x) of size wx

(together with all adjacent edges). Then it is still true that every vertex in the setf ′(X) :=

⋃

f(x)∈f(X) f′(x) sees more than half the vertices of Bad. So by double-

edge counting we see that on average, each vertex of Bad sees more than half of thevertices of f ′(X). Thus in the original graph, each vertex of Bad sees, on average,

a set f(Y ) with w(Y ) > |L|2. So there is at least one vertex in Bad actually seeing

such a set f(Y ), contrary to the definition of Bad.We shall now attempt to embed the remaining vertices of T ′ −L so that we will

be able to apply Hall’s Theorem to finish the embedding by embedding L. For this,we embed the remaining T ′ − L using all vertices of Bad, we proceed as follows.Embed the rest of T ′ − L in a greedy fashion, with the precaution that wheneverwe embed a vertex of T ′ − L, we immediately embed all of its leaf children. Also,we avoid v for the time being. As in the proof of Lemma 4.4, we see that we canchoose images for the vertices of C that see at least half of the unused vertices ofBad. Then we can embed half of the leaf children of each vertex x of C into verticesof Bad until we reach a vertex c ∈ C which has more children than there are unusedvertices of Bad. Since |L| > 2|Bad|, there is such a c. We embed c into v and fill

14

up the unused vertices of Bad with the leaf children of c. We continue greedily toembed all of T ′ − L. Let f(T ′ − L) be the image of T ′ − L.

Now, by Hall’s Theorem, to embed L in B \ f(T ′ − L), it is sufficient to provethat w(X ′) 6 |N(X ′) \ f(T ′ − L)| for all subset X ′ of X . Let X ′ be a subsetof X . If w(X ′) 6 |L|/2 then, since H has minimum degree at least (1 − 3ε)m′,N(X ′) \ f(T ′ − L)| > (1 − 3ε)m′ − m′ + |L| > (1

2− 7ε)m′ > (1

4+ 1

4ε)m′ > |L|/2.

If w(X ′) > |L|/2, then since Bad ⊆ f(T ′ − L), by definition of Bad, we haveN(X ′) \ f(T ′ − L) = B \ f(T ′ − L) and so |N(X ′) \ f(T ′ − L)| > |L|, because|B| > |D|. In both cases, w(X ′) 6 |N(X ′) \ f(T ′ − L)|. This completes the proof.

The ideas for the proofs of the remaining lemmas in this (and the subsequentsubsections) are substantially different (although we still use Hall’s theorem). Animportant tool is Lemma 4.7, which is needed for Lemma 4.8 below, and also forLemma 4.10 of Section 4.4.

For Lemma 4.7, let us introduce good orderings of parents. For a tree T and asubset L of its leaves, consider the set of parents P of L. Order the vertices of Pas p1, . . . , pm so that pi has at least as many leaf children as pi+1. Call any suchordering a good ordering of P .

Lemma 4.7 Let G be a graph with δ(G) >9m10, and let T be a tree with m edges

such that no vertex of T is incident to more than m6leaves. Let L be a subset of the

leaves of T such that |L| > 9m10. Suppose there is a good ordering p1, . . . , pa of the

parents P of L, and an embedding of T − L in G such that for each i 6 ⌊a/2⌋, wehave

|N(f(p2i−1)) ∪N(f(p2i))| > m. (8)

Then we can extend the embedding of T − L to an embedding T in G.

Proof. First of all, note that since no vertex has more than m6leaf children, for

any set S ⊆ P containing at most one of p2i−1, p2i, for each i 6 ⌊a/2⌋,

there are at mostm

6more leaves under S than under P − S, (9)

where we write ‘leaves under X ’ for leaves that are children of vertices in X .We use Hall’s theorem to show we can embed the vertices of L. For this, consider

the auxiliary bipartite graph H spanned between the set P ′ that arises from blowingup the image of each p ∈ P to a set Ap of size equal to the number of leaf childrenof p, and the set of unused vertices in G. For a ∈ Ap, the edge ab is present if p isadjacent to b.

A matching saturating P ′ shows we can complete the embedding, so assume thereis no such matching. By Hall’s theorem, there is a set S ′ ⊆ P ′ with |NH(S

′)| < |S ′|.Because of (8), S ′ can only contain vertices from one of Ap2i−1

, Ap2i , for each i 6

15

⌊a/2⌋, and so, by (9), we know that |S ′| 6 |P ′ − S ′| + m6. Thus |S ′| 6 2m

3. But, as

δ(G) > 9m10, and since for the embedding of T − L we used at most m

10vertices, it

follows that |N(S ′)| > |S ′|, a contradiction.

We continue with an analogue of Lemma 4.3 for bipartite host graphs. For itsproof, we will make use of Lemma 4.7.

Lemma 4.8 Let 0 < ε < 1200

, let H ′ = ((A,B), E) be a bipartite graph of minimumdegree at least (1 − ε)m′ such that A has at most ⌊(1 + ε)m′⌋ vertices and B hasexactly this many vertices. Let T ′ be a tree with m′ edges such that each vertex ofT ′ has at most m′

6leaf children. Then we can embed T ′ in H ′.

Proof. We let (C,D) be the unique 2-colouring of T ′ with |C| 6 |D|. Because ofthe minimum degree condition on H ′, we can greedily embed T ′ unless |C| < εm′+1,so we assume this is the case. Note that |C| > 2 as T ′ is not a star. Thus, we obtain|C| < 2εm′. We also obtain that the set L of leaves of T ′ in D has size at least(1− 2ε)m′ (for this, observe that rooting T ′ at a vertex of C, every non-leaf vertexin D has at least one child in C). Set T ′′ := T ′ − L. We will embed C in A andD in B. Consider a good ordering c1, . . . , ca of the parents of leaves in L. Wewant to embed T ′′ using an embedding f such that for every i 6 ⌊a

2⌋, we have

|NB(f(c2i−1)) ∪NB(f(c2i))| > m′. Then we are done with Lemma 4.7.As we embed T ′′, when we embed a vertex c = ci of C paired with a vertex

c′ = ci±1 which is already embedded, we choose as f(c) an unused vertex with thelargest number of neighbours in B −N(f(c′)). Let us next estimate how large thisnumber of neighbours will be.

Note that |N(f(c′))∩B| > (1− ε)m′ (by the minimum degree condition on H ′),and so, we have |B −N(f(c′))| 6 2εm′ (by our assumption on the size of B). Also,each vertex in B−N(f(c′)) misses at most 2εm′ vertices of A (again by the minimumdegree condition). Therefore, straightforward double-counting of non-edges betweenA and B − N(f(c′)) gives that there is a set A′ ⊆ A containing at least half thevertices of A such that each vertex in A′ misses at most

4ε2(m′)2

|A|/2 6 16ε2m′6 εm′

vertices of B − N(f(c′)). (For the first inequality, observe that m′ 6 2|A| becauseof the minimum degree condition.)

So, since we only embed |C| ≪ |A′| vertices in A, we will be able to choosean image f(c) that sees all but at most ⌊εm′⌋ vertices of B − N(f(c′)). Then,|NB(f(c)) ∪NB(f(c

′))| > m′ as desired. We thus find the desired embedding of T ′′

and hence of T .

The next lemma is an analogue of Lemma 4.5 for bipartite graphs.

16

Lemma 4.9 Let 0 < ε < 1200

and let H ′ = (A,B) be a bipartite subgraph of a graphG. Suppose H ′ has minimum degree at least (1 − ε)m′, both A and B contain atmost (1 + ε)m′ vertices, A contains a vertex v which has degree at least m′ in G,and every vertex of G−H ′ sees at least (1− 2ε)m′ vertices of G−H ′. Let T ′ be atree with m′ edges such that each vertex of T ′ has at most εm′

2leaf children. Then

we can embed T ′ in G.

Proof. We let (C,D) be the unique 2-colouring of T ′ with |C| 6 |D|. SetB′ := N(v) ∩ B and a := m′ − |B′|. Since we cannot embed the tree into (A,B′)using Lemma 4.6, we know that |B′| < |D|, and thus

|C| = m′ + 1− |D| 6 m′ − |B′| = a.

We embed a separator z for T ′ into v. We will embed the leaf children of z at the endof the process, which we can do because of our degree bound on v. Let K1, . . . , Kℓ

be the non-singleton components of T ′ − z. Every Ki contains a vertex of C, andthus ℓ 6 a.

Since z is a separator, we know that

∣

∣|D ∩ V (Ki)| − |C ∩ V (Ki)|∣

∣ 6m′ − 2

2(10)

for all i 6 ℓ. We let wi be the root of Ki, i.e. the vertex of Ki adjacent to z. Wewill embed the roots wi into neighbours of v in G and then embed the rest of thetree greedily in H ′.

First suppose that v has at least a neighbours in A. Successively embed theroots wi, in a way that ensures we can keep the embedding as balanced as possibleat each step. This means that when we are about to embed wi, we choose an imagefor wi in either A or B, so that the larger colour class of Ki will be forced to beembedded in that set among A, B that when we finish our embedding will containless of

⋃

j<i V (Kj). (If both A, B will contain the same number of vertices from⋃

j<i V (Kj), for instance when i = 1, we just arbitrarily choose either A or B forembedding wi.)

Next, embed greedily the remainder of the components Ki. This can be donesince the way we embedded the roots wi, together with (10), ensures that

∣

∣|D ∩⋃

j6i

V (Kj)| − |C ∩⋃

j6i

V (Kj)|∣

∣ 6m′ − 2

2

for each i 6 ℓ. Thus, throughout the embedding process of the Ki, we use at most3m′

4vertices on each side A, B.Now, if v has fewer than a neighbours in A, we attempt to perform the same

procedure. If we run out of neighbours of v in A during the embedding of the roots

17

wi, then we start to embed roots wi which were to be embedded into A intoN(v)−H ′

(this is possible as v has degree at least m′). We will embed the corresponding Ki

in G−H ′, using the large minimum degree of G−H ′. If at any point the total sizeof the components embedded in G−H ′ exceeds m′

4, then we stop embedding roots

wi in G−H ′. Instead, we embed the remaining wi in B and the remaining Ki in H ′

(this is possible because of the minimum degree of H ′). We will be able to embedthe components whose roots are embedded in G − H ′ because they have at most3m′

4vertices and this graph has minimum degree at least (1− 2ε)m′.

4.4 Graphs Without Very Dense Subgraphs

The main result of this section is Lemma 4.11. It says that if, in the situation ofTheorem 1.3, we cannot embed T in G, then either G is locally m-sparse (a situationwe dealt with in Subsection 4.1), or G contains at least one clique or bipartite (m, δ)-dense subgraph (see below for the definition). In the Subsections 4.2 and 4.3, wesaw how to use these subgraphs. Everything will be put together in the last part ofour proof, in Subsection 4.5.

Let us now define the subgraphs we are looking for. A subgraph H of G isclique (m,α)-dense if it has at most m + 1 vertices and minimum degree at least(1− α1/14)m. A connected bipartite subgraph H of G is bipartite (m,α)-dense if ithas minimum degree at least (1 − α1/14)m and each side of its (unique) bipartitionhas at most m vertices.

We first treat the case that T has many leaves. For this case, we need to makeuse of Lemma 4.7 from Subsection 4.3.

Lemma 4.10 For every sufficiently small α > 0 the following holds. Suppose G isa graph of minimum degree at least (1 − α)m with no clique (m,α)-dense subgraphand no bipartite (m,α)-dense subgraph, and let T be a tree with at most m edges.If T has at least (1− α1/7)m leaves, but no vertex of T is incident to more than m

6

leaves, then we can embed T in G.

Proof. Let L be the set of leaves of T and fix any good ordering p1, . . . , pa of theparents of L. We claim that we can embed all of T −L in G, via a good embeddingf , while maintaining that, for each i 6 ⌊a/2⌋, we have

|N(f(p2i−1)) ∪N(f(p2i))| > m. (11)

Then, Lemma 4.7 guarantees our partial embedding can be extended to an embed-ding of all of T . So we only need to prove we can find f satisfying (11).

For this, suppose that p = pj is the first vertex of T−L that cannot be embeddedwithout violating (11). Then there is an already embedded vertex p′ = pj±1 such thatthe pair p, p′ violates (11) for any embedding of p. Let q be the parent of p, and letA be a subset of size ⌈(1−α−α1/7)m⌉ of the unused neighbours of f(q). (Note that

18

there are that many unused neighbours of f(q) because |V (T − L)| 6 α1/7m+ 1 byassumption.) Let B be the set of (used and unused) neighbours of f(p′). Since (11)is violated for any embedding of p, we know that |B| 6 m and that

every vertex in A has degree less than m. (12)

Since δ(G) > (1−α)m, we have |B| > (1−α)m, and also, since (11) is violated,every vertex of A has at least (1−2α)m neighbours in B. So, there is a set B′ ⊆ B ofsize at least (1−

√2α)|B| such that each vertex in B′ has degree at least (1−

√2α)|A|

into A. Note that |B′| > (1−√2α)|B| > (1−

√2α)(1− α)m > (1− 2

√α)m.

Assume for a contradiction that A − B′ has size at most α1/7m. Then everyvertex of A ∩ B′ has degree at least

(1−√2α)|A| − α1/7m > (1−

√2α)(1− α− α1/7)m− α1/7m > (1− α1/14)m

in G[A ∩B′]. Hence G[A ∩ B′] is clique (m,α)-dense, a contradiction.Hence A−B′ has size at least α1/7m. Then A ∩ B′ = ∅, because the degree (in

G) of any vertex v ∈ A ∩ B′ would exceed

|A ∪ B′| − 2αm−√2α|A| > |B′|+ α1/7m− 2αm−

√2αm

> (1− 2√α)m+ α1/7m− 2αm−

√2αm

> m.

contradicting (12). So, the bipartite subgraph of G with sides A−B′ and B′−Ais bipartite (m,α)-dense, a contradiction. This proves the existence of an embeddingsatisfying (11), completing our proof.

We now use Lemma 4.10 together with Lemma 4.3 from the previous section toprove the main result of this section:

Lemma 4.11 For every sufficiently small positive constant α, and m > α−2, thefollowing holds for each tree T with at most m edges none of whose vertices hasmore than αm leaf children. If G is a graph of minimum degree at least (1 − α)mthat is not locally m-sparse and contains neither a clique (m,α)-dense subgraph nora bipartite (m,α)-dense subgraph then we can embed T in G.

Proof. If T has less than m− 1 edges, then the tree obtained from T by addinga path of length m − |V (T )| on any vertex of T also satisfies the hypothesis ofthe lemma. Thus, it suffices to prove the result for trees with m − 1 or m edges.Henceforth, we assume that T has m− 1 or m edges.

We choose α small enough to satisfy certain inequalities in the proof.By Lemma 4.10, we may assume that

T has fewer than (1− α1/7)m leaves. (13)

19

We let H be the densest subgraph of G with at most m + 1 vertices. We letδ := δ(H) be the minimum degree ofH , let a be its average degree and let w be someminimum degree vertex of H . Note that a >

m25, since G is not locally m-sparse. So,

as δ > a2(by our choice of H),

δ >m

50. (14)

Also,no vertex y outside of H sees more than δ + 1 vertices of H , (15)

as otherwise H − w + y contradicts our choice of H . Furthermore, we can assumethat

δ < (1− α1/14)m (16)

as otherwise H is a clique (m,α)-dense subgraph.We apply Observation 2.3 to obtain a vertex z such that the largest component of

T −z has fewer than m(1−α1/3) vertices and every other component has fewer thanα1/3m+1 vertices. We let F be a forest consisting of the union of some componentsof T − z with between α1/3m and 2α1/3m vertices. Note that since z has at mostαm leaf children (by the assumptions of the lemma), and since |F |−αm

2> αm

z has at least αm non-neighbours in F . (17)

We embed z into w and the neighbours of z in F into G−H ; this is possible becauseby (16) w has at least δ(G)− δ > (α1/14 −α)m > 2α1/3m neighbours in G−H . Weleave the remaining at least αm vertices of F to embed at the end of the process.

By (13), we know T − F has fewer than (1 − α1/7)m < (1 − 9α1/3)m leaves.Hence, by Lemma 2.2, we can choose a subtree T ′ of T − F containing z which has2⌈α1/3m⌉ + 2 vertices and a perfect matching.

As we are about to explain, we claim that either

(i) there are u, u′ ∈ V (H) such that dH(u) 6 δ + 3αm, and NH(u′) contains a set

A of ⌈δ − 4α1/3m⌉ vertices each of which sees at most δ + 7α1/3m vertices ofH at least δ − 4α1/3m of which are in NH(u), or

(ii) we can construct an embedding of T ′ so that for every x ∈ V (H) with dH(x) <δ+3αm, we have used at least 3αm vertices outside the closed neighbourhoodof x.

We will show that if (i) does not hold in H , then we can find an embeddingas in (ii). To do so, we root T ′ at z and consider a good iterative constructionprocess for T ′ into H in which (a) we embed the two vertices of each matching edgein consecutive iterations, and (b) we embed each vertex q in a randomly chosenunused element of N(f(p(q))). Using our lower bound of α−2 on m, we shall provethat with positive probability for every x ∈ V (H) with dH(x) < δ + 3αm, we haveused at least 3αm vertices outside the closed neighbourhood of x.

20

So consider a vertex x such that |NH(x)| < δ + 3αm. Let us first estimate theprobability that for a fixed matching edge e = {v1, v2} (of the perfect matching ofT ′) which does not contain z, we embed the second endpoint v2 of e outside NH(x).For this, we let A be the set of all vertices that are neighbours of the image ofp(v1) and see at most δ + 7α1/3m vertices of H at least δ − 4α1/3m of which arein NH(x). Since we assume (i) does not hold (for u = x and u′ = f(p(v1))), weknow that |A| < ⌈δ − 4α1/3m⌉, while there are at least δ − |V (T ′) − v1 − v2| >δ − 2⌈α1/3m⌉ available possible images for the first endpoint v1. Thus, irrespectiveof the embedding to this point, the probability that v1 is embedded in a vertexoutside A is at least 2α1/3. Therefore, again irrespective of the embedding to thispoint, the probability we embed v2 outside NH(x) is at least 4α

2/3. (For this, observethat every vertex outside A has at least 4α1/3m neighbours in H −NH(x) and thatat least 2α1/3m of them are unused.) We have shown1 that the number of non-neighbours of x used in the embedding is a random variable whose value dominatesBin(⌈α1/3m⌉, 4α2/3), where the binomial random variable Bin(n, p) is the sum of nindependent 0–1 random variables, each equal to 1 with probability p.

Thus the probability that there are less than 3αm such non-neighbours is boundedfrom above by the probability that Bin(⌈α1/3m⌉, 4α2/3) is less than 3αm. Chernoff’sBound (see [AS08, McD89]) states that for every t ∈ [0, np],

Pr (|Bin(n, p)− np| > t) < 2 exp

(

− t2

3np

)

.

Hence the probability that the number of non-neighbours of x used in the embeddingis less than 3αm is less than 2 exp(−αm/12).

Since the number of such vertices x (vertices with less than δ+3αm neighboursinH) is at mostm+1, the probability that there is a vertex x with |NH(x)| < δ+3αmsuch that less than 3αm non-neighbours of x are used in the embedding is at most(m + 1) × 2 exp(− 1

12αm) 6 2(m + 1) exp(− 1

12m1/2) because m > α−2. Since we

assumed m to be sufficiently large (since it is at least α−2), this is less than 1, andso there is an embedding as in (ii).

If we find an embedding as in (ii), then we can continue our good iterativeconstruction process on the rest of T − F , always embedding in a vertex of H ifpossible. Clearly, we embed at least δ+3αm+1 vertices in H . At this point, makinguse of (15), we can greedily embed F in the unused vertices of G−H .

So we will from now on assume that (i) holds. Then, we can find a subset Bof ⌈δ − 1.5α1/6m⌉ vertices of NH(u) each of which sees at least δ − 7α1/6m vertices

1We can decide for each matching edge e when we come to it, whether or not its second endpointis in NH(x), and then choose the embedding of its two endpoints conditional on our decision. Wecan make this decision by considering a random variable ze which is 1 with probability 4α2/3. Ifze = 1 we do not put the second endpoint of e in NH(x), otherwise we may or may not put thissecond endpoint in NH(X). The ze are independent.

21

of A. Indeed, otherwise there are at least

1.5α1/6m · (|A| − (δ − 7α1/6m)) > 1.5α1/6m · 3α1/6m > 4.5α1/3m2

non-edges between A and NH(u), but the way A was chosen allows for at most

|A| · (|NH(u)| − (δ − 4α1/3m)) 6 ⌈δ − 4α1/3m⌉ · (4α1/3m+ 3αm) < 4.5α1/3m2

such non-edges (here, we use that δ < m by (16) ). Clearly every vertex of A seesat least δ − 7α1/6m vertices of B.

Let us recapitulate the situation as follows. We found sets A,B ⊆ V (H) suchthat

|A| = ⌈δ − 4α1/3m⌉, |B| = ⌈δ − 1.5α1/6m⌉ (18)

and

the minimum degree from A to B and from B to A is at least δ − 7α1/6m. (19)

Case 1: A−B and B − A both have size at least 25α1/6m.

Let [A − B,B − A] denote the bipartite subgraph of G spanned by the edgesbetween A− B and B −A. Then

[A− B,B − A] has minimum degree at least 17α1/6m. (20)

Furthermore, each vertex of A∩B sees at least |B| − 7α1/6m+ |A−B| − 7α1/6m >

|B|+ 11α1/6m vertices of A ∪ B, and thus,

each vertex of A ∩B sees at least δ + 9α1/6m vertices of A ∪B. (21)

By (13), T −F has fewer than (1−α1/7)m leaves, and by definition |T −F |2α1/3.Hence, T−F has fewer than |T−F |−33α1/6m leaves. So, by Lemma 2.2, we can finda subtree T ∗ of T − F with 2⌈16α1/6m⌉ vertices which contains z and has a perfectmatching and hence a 2-colouring with colour classes of equal size. Using (20), weembed T ∗ into [A− B,B − A], with z in A− B.

We claim that at this point, for every vertex x of A∪B with less than δ+α1/6mneighbours in A ∪ B,

we have embedded at least 8α1/6m vertices in non-neighbours of x. (22)

For this, it suffices to observe that x /∈ A ∩ B by (21), and if x ∈ A, say, then weembedded at least 16α1/6m vertices in A−B, but x only sees at most δ + α1/6m−dB(x) 6 8α1/6m of these. (Here we used (19) for the bound on dB(x).)

We continue embedding T − F into H [A ∪ B] until we have embedded at leastδ+α1/6m+1 vertices into it, which we can do because of (19) and (22). By definition

22

of F , z has at most 2α1/3m neighbours in F . We can embed these into G− V (H),since f(z) has at least

δ(G)− δ > (1− α)m− (1− α1/14)m > 2α1/3m

neighbours outside H (we used (16) for the first inequality). We can then completegreedily the embedding of T − F as at least αm vertices of F have not yet beenembedded by (17). Finally, complete the embedding of F in G − V (H); this ispossible because in G− V (H) every vertex has degree at least (1− α)m− δ − 1 by(15), and at most m− δ − α1/6m vertices of T are embedded in G− V (H).

Case 2: One of A−B or B −A has size at most 25α1/6m.

Since by (19), each vertex of A misses at most 7α1/6m vertices of B, and viceversa, G[A∩B] has minimum degree at least |A∩B|−7α1/6m. We consider a largestinduced subgraph H ′ of G with at most m+ 1 vertices and at most 7α1/6m|V (H ′)|non-adjacent pairs of vertices, chosen so as to maximize the number of edges in H ′.So, if H ′ has minimum degree δ′ then

every vertex outside H ′ has degree at most δ′ + 1 in H ′. (23)

Note that since G[A ∩ B] is one possible choice for H ′,

|V (H ′)| > |A ∩B| > min{|A|, |B|} − 25α1/6m > δ − 27α1/6m >m

100, (24)

where we used (18) in the second-to-last inequality and (14) in the last one. Weobtain a subgraph H∗ of H ′ by iteratively deleting vertices which are non-adjacentto more than α1/13m

3vertices in the current subgraph. Then the minimum degree

m∗ of H∗ is bounded by

δ∗ > |V (H∗)| − α1/13m

3. (25)

Clearly we delete at most α1/13m10

vertices, that is,

|V (H ′)| − |V (H∗)| 6 α1/13m

10. (26)

If |V (H∗)| exceeds (1 − α1/13)m then as H∗ has minimum degree at least δ∗ >

|V (H∗)| − α1/13m3

> (1 − α1/14)m, we obtain that H∗ is an (m,α)-dense clique,contradicting our assumption that no such exist. So we can assume that

|V (H∗)| 6 (1− α1/13)m. (27)

Observation 2.3 implies we can choose a vertex z∗ of T such that the largestcomponent of T−z∗ contains at most (1−α1/13

2)m vertices and every other component

23

of T − z∗ contains fewer than α1/13m2

vertices. We choose a smallest possible forest

F ∗ consisting of the union of components of T−z∗ whose total size is between α1/13m2

and α1/13m. We note that since z∗ is incident to at most αm leaves,

F ∗ contains at leastα1/13m

6non-neighbours of z∗. (28)

First suppose δ′ (the minimum degree of H ′) is at most |V (H ′)| − 1 − 2α1/13m3

.We use a good iterative construction process to embed T − F ∗ into G with z∗ ina vertex of H∗ and using vertices of H∗ when possible. By (25), we use at least

|V (H∗)| − α1/13m3

+ 1 vertices of H∗ before embedding any of T − F ∗ outside H∗.When we are about to first embed a vertex outside of H∗, we proceed as follows.

We start by embedding the neighbours of z∗ in F ∗ into G − V (H ′). Observethat this can be done, since because of (28), we know that z∗ has at most 5

6α1/13m

neighbours in F ∗, while f(z∗) has at least

(1− α)m− |V (H∗)| − |V (H ′ −H∗)| > (9

10α1/13 − α)m

neighbours in G−H ′ (here, we used (26) and (27)). Then we finish our embeddingof T − F ∗, just using the minimum degree of G. Finally, we embed the rest of F ∗

in G− V (H ′), using (23), our assumption on δ′, and the fact that we used at least

|V (H∗)| − α1/13m3

+ 1 vertices of H∗.

So we can assume that δ′ > |V (H ′)| − 1 − 2α1/13m3

. Since H ′ is not an (m,α)-

dense clique, it follows that |V (H ′)| 6 m(1 − α1/14

2). We choose (the unique value

of) ε such that δ′ = (1 − 2ε)(|V (H ′)| − 1). Choose a subtree T ′ of T − F ∗ withm′ = (1 − ε)(|V (H ′)| − 1) edges that contains z∗ and subject to this has as fewleaves as possible. We note that this implies if a vertex of T ′ has two leaf childrenthen all its leaf children are also leaves of T − F ∗.

If no vertex of T ′ is incident to more than εm′

2leaves then Lemma 4.3, with

ε := (α1/13

3)( m

|V (H′)|−1), ensures that we can embed T ′ in H ′ with z∗ embedded in a

vertex of minimum degree in H ′. Note that for the application of Lemma 4.3, weuse that because of (24) and the fact that we can make m as large as we want bymaking α small, we know that m

|V (H′)|−1is at most 101, ensuring that ε is sufficiently

small. When we stop there are at most εm′ unused vertices of H ′.If some vertex of T ′ is incident to more than εm′

2leaf children then all but one of

these leaves are also leaves of T . So, by hypothesis, εm′

2< αm+ 2. In this case,we

just use a good iterative construction process to embed as much of T ′ into H ′ aspossible where to begin we embed z∗ in a minimum degree vertex of H ′. When westop there are at most 2εm′ < 4αm+ 8 unused vertices of H ′.

In either case, as above, we then embed all of the neighbours of z∗ in F ∗ intoG − H ′ which we can do because of our upper bound on the size of |V (H)|. We

24

then finish our embedding of T − F ∗, just using the minimum degree of G. For theembedding of the rest of F , it is enough to observe that by our chocie of H ′, everyvertex of V (G) − V (H ′) misses at least max(7α1/6m, 2εm′ − 1) vertices of H ′ andthe number of unused vertices of H ′ is at most max(εm′, αm+ 8).

4.5 Finishing Things Off

In this section we prove Theorem 1.3. We choose α < 1/20015 sufficiently small sothat Lemma 4.11 holds, and that other inequalities implicitly given in this sectionhold. We choose γ = α2. Note that we can assume m >

1γ= 1

α2 as otherwisethe graph has minimum degree greater than m − 1 so at least m, and we can justgreedily embed T . We can also assume that no vertex has γm or more leaf childrenas otherwise we can embed this vertex in a maximum degree vertex, greedily embedthe tree except for its leaf children and then greedily embed these children.

By Lemmas 4.1 and 4.11, we may assume G contains a clique or bipartite (m,α)-dense subgraph. For a clique (m,α)-dense subgraph D of G, by an expansion of Dwe mean a graphH obtained by iterately adding vertices (one at a time) of G−V (D)which see at least (1 − α1/15)m vertices of the current expansion. For a bipartite(m,α)-dense subgraph D = (A,B) of G, by an expansion of D we mean a graphH = (A′, B′) obtained by iterately adding one at a time vertices v of G − V (D)which see at least (1−α1/15)m vertices of one of the sides of the current expansion;we then add v to the other side, and forget about all edges from v to this side. Amaximal expansion of D is an expansion H as defined above of maximal size.

G contains an expansion H of a clique (m,α)-dense subgraph with |V (H)| =1+⌈(1−α1/15)−1m⌉, then we can embed T within it, by Lemma 4.3, with ε := α1/15 <1

200. (For this, observe that the minimum degree ofH is at least ⌈(1−α1/15)m⌉ > (1−

2ε)(|V (H)|−1), while the number of edges of the tree T is m 6 (|V (H)|−1)(1−ε).)So we can assume for all expansions H of clique (m,α)-dense subgraphs of G wehave

|V (H)| < 1 + (1− α1/15)−1m. (29)

Similarly, if G contains an expansion H = ((A′, B′), E ′) of a bipartite (m,α)-dense subgraph D = ((A,B), E) with max{|A′|, |B′|} = ⌊(1 + α1/15)m⌋ then wecan embed T within it, by Lemma 4.8. So we can assume for each expansionH = ((A′, B′), E ′) of every bipartite (m,α)-dense subgraph of G we have

max{|A′|, |B′|} 6 (1 + α1/15)m. (30)

We will show below that if we cannot embed T , then for each maximal expansionH of a clique or bipartite (m,α)-dense subgraph of G, it holds that

(A) no vertex of G−H sees more than 2γm vertices of H , and

(B) no vertex of H sees more than 2γm vertices of G−H .

25

Now, assuming (A) and (B) hold, we consider a maximal sequence D1, . . . , Dℓ

of clique and bipartite (m,α)-dense subgraphs of G, together with correspondingmaximal expansions H1, . . . , Hℓ. More precisely, we chooseDi as a clique or bipartite(m,α)-dense subgraph of G − ⋃

j<iHj, and let Hi be its maximal expansion inG − ⋃

j<iHj. Note that Di is clique or bipartite (m,α)-dense in G and by (B),applied to the graphs Hj with j < i, we know that Hi is also a maximal expansionof Di in G.

We will show below that moreover, if we cannot embed T , then

(C) no vertex of V (G)−⋃ℓi=1Hi sees more than 10γm vertices of

⋃ℓi=1Hi.

Thus, if G−⋃ℓi=1Hi is non-empty then by Lemmas 4.1 and 4.11, we can embed

T within it. So, choosing a vertex v ∈ V (G) of maximal degree, we can assume thatv is contained in one of the Hi.

Now if v is in the expansion of a bipartite (m,α)-dense subgraph, then Lemma 4.9,together with (30) and (A), tells us that we can embed T . So we can assume v isin the expansion H of a clique (m,α)-dense subgraph. If |V (H)| 6 1 + (1 + 3γ)m,then Lemma 4.5, together with (A) and (B), gives an embedding of T in G. So

|V (H)| > 1 + (1 + 3γ)m. Setting ε := |V (H)|−1−m|V (H)|−1

we see that (B) guaran-

tees that the minimum degree of H is at least (1 − γ)m − 2γm = (1 − 3γ)m >

(1− 2ε)(|V (H)| − 1). Furthermore our upper bound (29) on the size of expansionsensures ε 6 α1/15 < 1

200. Finally our lower bound on V (H), ensures that for suffi-

ciently small γ, ε = 1 − m|V (H)|−1

> 1 − 11+3γ

= 3γ1+3γ

> 2γ. Hence we can embed Tusing Lemma 4.3.

This completes the proof of the theorem. It only remains to show (A), (B)and (C).

To prove (A) we consider the expansion H of some clique or bipartite (m,α)-dense subgraph D of G. Note that by the definition of an expansion, G − H hasminimum degree at least (α1/15−γ)m. Let w be a vertex outside ofH with maximumdegree into H . Let d be the number of its neighbours in H and assume for acontradiction that d > 2γm.

Our plan is to find an embedding of T in G, here is an outline of the proof.We distinguish between two cases: First, we treat the case that d is relatively large(almost m/2 or larger). In this case we embed a suitable vertex z of T in w, a fewsmall components of T − z outside of H , and the main part of T in H . The othercase is that d is rather small (between 2γm and almost m/2). In that case, weembed a suitable vertex z of T in w, and embed into H a set C of components ofT − z whose union contains a little bit more than d, namely d+ γm, vertices. Thisis possible since T − z has at most γm singleton components, and so the number ofneighbours of z in

⋃ C is at most d. We then embed the rest of T outside H . Letus now turn to the details of this plan.

26

Case 1: d > (12− α1/15

6)m.

In this case we choose a vertex z of T such that the largest component of T − zhas at most (1 − α1/15

3)m + 1 vertices and every other component has fewer than

α1/15m3

vertices (this is possible by Observation 2.3). We embed z into w. We choosesome components including all the (at most γm) singleton components, so that the

union of these components has between α1/15m3

and 2α1/15m3

vertices. We embed thesecomponents greedily into G−H . Since the remaining components of T−z each haveat least two vertices, there are at most 1

2(1 − α1/15

3)m < d of them. We embed the

roots (neighbours of z) of these components into neighbours of w in H , preferringvertices of D.

We then proceed to embed greedily into H all those components of T − z whoseroot was embedded in H − D. If such components exist, then, since there are atmost 4α1/15m vertices in H −D (at most ((1 − α1/15)−1 + α1/14)m 6 3α1/15m if Dis clique dense by (29), and at most 4α1/15m if D is bipartite dense by (30)), andsince we preferred vertices of D for putting down the roots, we must have embeddedat least d − 4α1/15m >

m3roots of other components into D. So, as we already got

rid of singleton components, there are at least m3vertices in components whose root

is in D which we are not yet embedded. Thus, the minimum degree of H , whichis (1 − α1/15)m, is sufficient for embedding all components with roots in H − D.Finally, we embed all those components whose root was embedded in D. For this,observe that being a (m,α)-dense subgraph, D has minimum degree (1 − α1/14)m,

which is sufficient for embedding the rest of T (since at least α1/15m3

vertices of Twere already embedded outside H).

Case 2: 2γm < d 6 (12− α1/15

6)m.

Then G−H has minimum degree

δ(G−H) > m− d− γm. (31)

We choose a vertex z of T such that the largest component Cmax of T − z hasfewer than 1+m− d− γm vertices and every other component has at most d+ γmvertices (possible by Observation 2.3), and embed z into w. Let CT−z be the set ofcomponents of T − z. Take a smallest set C ⊆ CT−z − {Cmax} with

d+ γm 6 |⋃

C∈C

V (C)|. (32)

Clearly,

|⋃

C∈C

V (C)| 6 2d+ 2γm. (33)

We claim that moreover,

if C has singleton components, then |⋃

C∈C

V (C)| 6 2d− ⌊γm⌋. (34)

27

In order to see (34) suppose that |⋃C∈C V (C)| > 2d−⌊γm⌋+1. We need to show thatC has no singleton components. For this, it suffices to observe that by the minimalityof C, for each component C∗ ∈ C we have that |⋃C∈C,C 6=C∗ V (C)| 6 ⌈d + γm⌉ − 1.So

|V (C∗)| > 2d− ⌊γm⌋ + 1− (⌈d+ γm⌉ − 1) = d− ⌊γm⌋ − ⌈γm⌉ + 2 > 1,

where for the last inequality we apply our hypothesis that d > 2γm.Next, we wish to show that

|C| 6 d. (35)

If C has singleton components, then there are at most ⌊γm⌋ such components,and (35) follows from the fact that by (34),

|C| 6 ⌊γm⌋ + |⋃C∈C V (C)| − ⌊γm⌋2

6 ⌊γm⌋ + 2d− 2⌊γm⌋2

= d.

If C has no singleton components, and additionally, |⋃C∈C V (C)| 6 2d, then |C| 6|⋃

C∈CV (C)|

26 d, as desired, so let us now assume that |⋃C∈C V (C)| > 2d. Then |C| 6

3, as otherwise the set C′ obtained from C by deleting the smallest component satifies|⋃C∈C′ V (C)| > 3

4· 2d > d + γm (since d > 2γm), contradicting the minimality of

C. Moreover, since d > 2γm > 2, we know that d > 3. Thus again, |C| 6 d. Thiscompletes the proof of (35).

We now embed T − z. By (31) and by (32), the minimum degree of G − H islarge enough to greedily embed into G − H all the components of T − z that arenot in C. Next, we embed the (by (35) at most d) roots of the components from Cinto H , as above preferring vertices in D over vertices in H − D. We then embedall components whose root was put into H −D, and finally embed the componentswith root embedded in D. In order to see that we succeed in embedding all of T ,we argue similarly as in the previous case: For the components with root in H −D,note that again, we must have embedded at least d−4α1/15m roots of non-singletoncomponents into D, so, unless d − 4α1/15m < α1/15m, we can argue as above thatthe minimum degree of H is sufficient. On the other hand, if d−4α1/15m < α1/15m,that is, if d < 5α1/15m, then by (33),

|⋃

C∈C

V (C)| 6 2d+ 2γm < 10α1/15m+ 2γm,

so again, the minimum degree of H is sufficient. For the components with root inD, note that as above, the minimum degree of D is sufficient for embedding thembecause by (33) at least

m− |⋃

C∈C

V (C)| > m− 2d− 2γm >α1/15m

3− 2γm > α1/14m

28

vertices of T were already embedded outside H . This completes the proof of Case2, and thus of (A).

To prove (B) we consider the expansion H of some clique or bipartite (m,α)-dense subgraph D of G. We let w be a vertex of H which has maximum degreedG−H(w) outside of H and set

d := min{dG−H(w), α1/15m}.

Then, H has minimum degree at least m − d − γm (this is clear if d = dG−H(w),and follows from the fact that H is an expansion in the case that d = α1/15m). Weembed a separator z for T into w. We choose a minimal set C of components ofT −z containing at least d+γm vertices. Since at most γm of these components aresingletons, we need at most d

2+ γm < d components. Furthermore, their total

size is at most m2

(as the size of the components of T − z is bounded by thisnumber, since z is a separator). The minimum degree of H is clearly enough togreedily embed into H all those components of T − z that are not in C. We thenembed the components from C into G −H . We embed the (at most d) neighboursof z first. After that, the minimum degree of G − H (which, by (A), is at least(1 − γ − 2γ)m >

m2>

⋃

C∈C |V (C)|) ensures we can embed the remainder of thecomponents from C. This completes the proof of (B).

It remains to prove (C). We shall do so by inductively proving that if we cannotembed T , then for every j between 1 and ℓ,

(C’) no vertex of V (G)−⋃ji=1Hi sees more than 10γm vertices of

⋃ji=1Hi.

For j = 1, (C’) holds by (A). Assuming (C’) holds for j − 1, let us showthat (C’) also holds for j. By (A), no vertex of V (G) − ⋃j

i=1Hi sees more than

10γm+ 2γm = 12γm vertices of⋃j

i=1Hi. Thus, G−⋃ji=1Hi has minimum degree

at least m− 13γm.Suppose there is a vertex w ∈ V (G)−⋃j

i=1Hi which sees at least 10γm vertices

in⋃j

i=1Hi. Our aim to show that we can then embed T . We embed a separator zfor T into w. We choose a minimal set C of components of T − z containing at least13γm vertices. Since at most γm of the components in C are singletons, we knowthat |C| 6 7γm. Furthermore,

⋃

C∈C |V (C)| 6 m2. We then embed the components

of T − z that are not in C greedily into G − ⋃ji=1Hi, using the minimum degree

of G−⋃ji=1Hi. Finally, we embed the components from C into

⋃ji=1Hi, embedding

the (at most 7γm) neighbours of z first, and using the minimum degree of the Hi

for the rest of these components (this works since we embedded at least m2vertices

outside of⋃j

i=1Hi). This shows (C’), and thus completes the proof of (C).

29

Acknowledgements

The authors would like to thank the referees for their comments which greatlyimproved the readability and quality of the paper.

References

[AKS95] M. Ajtai, J. Komlos, and E. Szemeredi. On a conjecture of Loebl. InGraph theory, combinatorics, and algorithms, Vol. 1, 2 (Kalamazoo,MI, 1992), 1135–1146. Wiley, New York, 1995.

[AS08] N. Alon and J. H. Spencer. The probabilistic method. Wiley, 2008.

[BPMS19] G. Besomi, M. Pavez-Signe, and M. Stein. Degree conditions for em-bedding trees. SIAM J. Discrete Math., 33:1521–1555, 2019.

[BD96] S. Brandt and E. Dobson. The Erdos–Sos conjecture for graphs of girth5. Discr. Math., 150:411–414, 1996.

[Coo09] O. Cooley. Proof of the Loebl-Komlos-Sos conjecture for large, densegraphs. Discr. Math., 309(21):6190–6228, 2009.

[EFLS95] P. Erdos, Z. Furedi, M. Loebl, and V. T. Sos. Discrepancy of trees.Studia Sci. Math. Hungar., 30(1-2):47–57, 1995.

[Hax01] P. E. Haxell. Tree embeddings. J. Graph Theory, 36(3):121–130, 2001.

[HKP+17a] J. Hladky, J. Komlos, D. Piguet, M. Simonovits, M. Stein, and E. Sze-meredi. The approximate Loebl–Komlos–Sos Conjecture I: The sparsedecomposition. SIAM J. Disc. Math. 31(2):945–982, 2017.

[HKP+17b] J. Hladky, J. Komlos, D. Piguet, M. Simonovits, M. Stein, and E. Sze-meredi. The approximate Loebl–Komlos–Sos Conjecture II: The roughstructure of LKS graphs. SIAM J. Disc. Math. 31(2):983–1016, 2017.

[HKP+17c] J. Hladky, J. Komlos, D. Piguet, M. Simonovits, M. Stein, and E. Sze-meredi. The approximate Loebl–Komlos–Sos Conjecture III: The finerstructure of LKS graphs. SIAM J. Disc. Math. 31(2):1017–1071, 2017.

[HKP+17d] J. Hladky, J. Komlos, D. Piguet, M. Simonovits, M. Stein, and E. Sze-meredi. The approximate Loebl–Komlos–Sos Conjecture IV: Embed-ding techniques and the proof of the main result. SIAM J. Disc. Math.31(2):1072–1148, 2017.

[HP15] J. Hladky and D. Piguet. Loebl–Komlos–Sos Conjecture: dense case.J. Combin. Theory Ser. B, 116:123–190, 2016.

30

[HPS+15] J. Hladky, D. Piguet, M. Simonovits, M. Stein, and E. Szemeredi.The approximate Loebl–Komlos–Sos conjecture and embedding treesin sparse graphs. Electron. Res. Ann. Math. Sci., 22:1–11, 2015.

[McD89] C. McDiarmid. On the method of bounded differences. In Surveysin combinatorics, volume 141 of London Math. Soc. Lecture Note Ser.,pages 148–188. Cambridge Univ. Press, 1989.

[PS12] D. Piguet and M. J. Stein. An approximate version of the Loebl-Komlos-Sos conjecture. J. Combin. Theory Ser. B, 102(1):102–125,2012.

[RS19a] B. Reed and M. Stein. Spanning trees in graphs of high minimumdegree with a universal vertex I: An approximate asymptotic result.arXiv:1905.09801.

[RS19b] B. Reed and M. Stein. Spanning trees in graphs of high minimum degreewhich have a universal vertex II: A tight result. arXiv:1905.09806.

[SW97] J.-F. Sacle and M. Wozniak. A note on the Erdos–Sos conjecture forgraphs without C4. J. Combin. Theory Ser. B, 70(2):229–234, 1997.

[Zha11] Y. Zhao. Proof of the (n/2−n/2−n/2) conjecture for large n. Electron.J. Combin., 18(1):P27, 2011.

31